Analyzing massive code bases has become a staple of modern software engineering research. This has happened as a welcome side-effect of the advent of public large-scale software repositories such as GitHub. Yet, finding which projects to analyze is a labor-intensive process that can lead to biased analysis results if the selection is not representative. The search interfaces exposed by mainstream software repositories do not allow researchers to formulate anything but very basic queries. This paper reports on Code DJ , an infrastructure designed to assist researchers in querying such repositories and identifying projects of interest. The infrastructure is composed of two subsystems: a persistent datastore that is constantly updated with information acquired from its target large-scale software repository (in our case GitHub), and an in-memory database with a query interface written in Rust and designed to follow popular data science API principles. Our infrastructure has built-in support for reproducibility. Users can formulate historical queries that are answered deterministically using historical states of the datastore; thus researchers can always reproduce published results. To illustrate the benefits of the proposed system, we revisit a paper aiming to establish a correlation between programming languages and software defect. Using Code DJ , we identify biases in the dataset used in the original paper. By repeating the analysis performed by the original authors with new data, we demonstrate that the results of the paper are highly sensitive to the choice of projects.